You are on page 1of 53

CNS, JC talks

Exact mean-field models for


spiking neural networks with
adaptation
Liang Chen, Sue Ann Campbell
Journal of Computational Neuroscience (2022) 50:445–469
https://doi.org/10.1007/s10827-022-00825-9

Presented by Gustavo Patow, ViRVIG–UdG & CNS


Why?
Personal reasons…
The Izhikevich model
E.M. Izhikevich, 2003
https://www.izhikevich.org/publications/spikes.htm

u = 0.04 × u + 5 × u + 140 - w + I
2

w = a × (b × u - w)
if u > 30mV threshold
u=c
w=w+d
Variables: a, b, c, d
6
u = 0.04 × u 2 + 5 × u + 140 - w + I
w = a × (b × u - w)
if u > 30mV u
u=c
w
w=w+d

7
With code!
Equivalence


θ

vk(t)
-∞

wk(t)

QIF model vk (t) = tan (


θk (t)
) Theta model
2
+ Adap. + Adap.

Izhikevich 2003, IEEE Trans. Neural Networks, 14(6) Adapted from Wiki
Mean-Field formulation
The network of Izhikevich neurons

All-to-all coupled
vk(t)

wk(t)
CA3 Izhikevich neurons…
They use a variant of the standard Izhikevich’s neuron:

vk’ = vk(vk − 𝛼) − wk + 𝜂k + Iext + Isyn,k


wk’ = a(b vk − wk)
If vk ≥vpeak, then vk ← vreset
and wk ← wk + wjump

Note: dimensionless version!


[Dur-e-Ahmad et al., 2012]
CA3 Izhikevich neurons…
Neurons are connected by the standard synaptic current model

Isyn,k = gsyn sk(er− vk)

• where (assumed to be the same for all neurons)


• er is the reversal potential
• gsyn the maximum synaptic conductance
• The synaptic gating variable, sk ∈ [0, 1], represents the proportion of
ion channels open in the postsynaptic neuron as the result of the
firing in presynaptic neurons
• For a network with all-to-all connectivity, sk is homogeneous across the network
as every post synaptic neuron receives the same summed input from all the
presynaptic neurons,
• Thus sk = s
Gating Variable(s)
• The mechanism of synaptic transmission can be formally
described by a linear system of ODEs with a sum of delta
pulses corresponding to the times a neuron fires a spike
[Ermentrout & Terman, 2010]
• The single exponential synapse is modelled by

sʹ=−s + sjump ∑k=1∑tjk<t𝛿(t−tjk)


𝜏s N
Aside: Lorentzian approximation (next)
Population density approach
• Excellent readings:
• Felix Apfaltrer, Cheng Ly & Daniel Tranchina (2006) Population density
methods for stochastic neurons with realistic synaptic kinetics: Firing
rate dynamics and fast computational methods, Network:
Computation in Neural Systems, 17:4, 373-418,
DOI:10.1080/09548980601069787
• Nykamp, D.Q., and Tranchina, D. (2000) A population density
approach that facilitates large-scale modelling of neural networks:
analysis and an application to orientation tuning. J. Comput.
Neurosci. 8, 19–50. doi:10.1023/A:1008912914816
Mean-field modelling ideas ηk

dr/dt = ? d⟨ v⟩ /dt = ?
• Lorentzian distribution ansatz
• [Montbrió et al. 2015, Phys. Rev. X 5]
• [Ott and Antonsen, 2008, Chaos, 18(3)]

Population density approach

x(η, t) Δη

y(η, t) η¯

ρ(v, w, η, t) L(η)
Mean-field modelling ideas ηk

d⟨ w⟩ /dt = ?
Moment closure assumption:
• [Nicola & Campbell, 2013b]: The validity of this assumption at
high firing rates is supported by numerical simulations of the
full network

⟨ w | v, η⟩ ≈ ⟨ w | η⟩
vk(t)

wk(t)
Mean-field modelling ideas ηk

d⟨ w⟩ /dt = ?
• Perturbation theory:
the mean adaptation with the parameter η is sufficiently
greater than the after-spike jump size [Nicola & Campbell.
2013, JCN, 35(1)]
⟨ w | η⟩ ≫ w jump

vk(t)

wk(t)
The mean-field model
ODEs:

rʹ = Δη/π + 2r⟨v⟩ − (α + gsyn s)r


⟨v⟩ʹ = ⟨v⟩2 − α⟨v⟩ − ⟨w⟩ + η + Iext + gsyn s(er − ⟨v⟩) − π2r2
⟨w⟩ʹ = a (b⟨v⟩ − ⟨w⟩) + wjump r
sʹ = − s/τs + sjump r
Network vs Mean-field Model

Asynchronous tonic firing (EPs) Data: pyramidal neurons in CA3 of hippocampus


Synchronous bursting (POs)
[Dur-e-Ahman et al. 2012, JCN, 33(1)]
[Hemond et al. 2008, Hippocampus, 18(4)]
Network vs Mean-field Model 2
0.35

0.3

0.25

0.2

0.15

0.1
EP-
EP+
0.05
PO-
PO+
0
0.05 0.1 0.15 0.2 0.25 0.3

Asynchronous tonic firing (EPs)


Synchronous bursting (POs)
Network of two-population Izhikevich
neurons
Strongly adapting

Weakly adapting

Data: pyramidal neurons in CA3 of hippocampus


vk(t)
[Dur-e-Ahman et al. 2012, JCN, 33(1)]
[Hemond et al. 2008, Hippocampus, 18(4)]
wk(t)

[Izhikevich 2003, IEEE Trans. Neural Networks, 14(6)]


Network of two-population Izhikevich
neurons
Strongly adapting (p) Weakly adapting (q)
Network of two-population Izhikevich
neurons
• Populations: Strong (p) / Weak (q)
• With currents and gating variables:

• K is the proportion k=Np/(Np+Nq)


Two-pop. Network vs Mean-Field Model
SA

WA

7
Two-pop.
What can we learn from the model ?
New discovery

SA

WA

Explain phenomena observed in


experimental studies
• [Dur-e-Ahman et al. 2012, JCN, 33(1)]
• [Hemond et al. 2008, Hippocampus, 18(4)]
New discovery!!

8
Assumptions…
To assess the validity of the mean-field approximation, we examine all the
assumptions that are imposed during the derivation
Assumptions (1)
• All-to-all connectivity within the population and
between different populations
• Reasonable for the application to CA3 region of
hippocampus
• There are formalisms for sparse networks [Ferguson
et al., 2015; Di Volo & Torcini, 2018; Biet al., 2021;
Lin et al., 2020]
Assumptions (2)
• N → ∞ , the
thermodynamic
limit
• As the number of
neurons increases,
the spread of the
network variables
around the mean
narrows and gets
closer to the
dynamics of the
mean-field model
Assumptions (3)
• ⟨w|𝜂⟩ ≫ wjump , the mean adaptation variable with the
parameter 𝜂 is sufficiently greater than the homogeneous
after-spike jump value
• This assumption is required for the differential equation of ⟨w⟩
• However, the mean-field description still captures the essential shape
and frequency of the firing activity of the network
• The accuracy could be improved by inclusion of high-order terms in
the Taylor expansion ⇒ extra term for ⟨w⟩’
Assumptions (4)
• ⟨w|v,𝜂⟩ = ⟨w|𝜂⟩, first-order moment closure
assumption, also called the adiabatic
approximation
• This assumption entails fast dynamics of the
membrane potential
• We could employ a high-order moment closure
approximation,
• Although we need to assess the cost of the added effort
in terms of the improvement of the accuracy of the
resulting mean-field model [Ly & Tranchina, 2007]
Assumptions (5)
• The Lorentzian ansatz on the conditional density
function

• A crucial step for the population density function


• [Nicola and Campbell, 2013b] show how changing the
expansion of the population density function can drastically
change the resulting mean-field model
Assumptions (6)
• vpeak = −vreset → ∞, limit of the resetting rule when neurons fire
• Used parameter values are based on actual neuronal data except the
resetting values
• Essential for the validity of the Lorentzian ansatz
• And for linking the Izhikevich model to the Theta model
• When dealing with a biological network based on experimental data,
changing vpeak and vreset can affect firing rates and estimation of <v>
• In numerical experiments: vpeak = −vreset = 200
• Could be addressed by adding a refractory period to the network
model [Montbrio et al., 2015],
• Makes the firing rate of the asynchronous tonic firing (EPs) and <v> of the
network match those of the theta model and hence the mean-field model
• But lacks adaptation and thus, cannot have synchronous bursting (POs)
Assumptions (7)
• Lorentzian distribution of the heterogeneous current

• Many parameters can be the sources of heterogeneity in a network


(e.g., η but also gsyn)
• Sharply reduces the complexity of the final mean field model
• The Lorentzian distribution is physically implausible since both its
expected value and its variance are undefined
• Could use a Gaussian [Klinshov et al., 2021], but would result in a
much more complex expression…
Assumptions (8)
• 𝜂 ∈ (−∞, ∞), range of the heterogeneous current
• This assumption is adopted in evaluation of the
integrals using the residue theorem in the deduction
process
• For the neural network to be realistic in spite of this
requirement, the distribution range of the
heterogeneous parameter should be much wider
than its half-width at half maximum
Code available!
• Matlab
• All figures
• Well organized
• Could be streamlined, though…
• In general, OK!
Summary
• Derivation of exact mean-field models for Neural
network with spike adaptation (Izhikevich neurons)
• One population
• Two-coupled populations
• Great agreement between the network and mean-field
models
• Results from bifurcation analysis of the mean-field
model
• Impact of proportion of strongly adapting neurons
• Impact of heterogeneity of firing rates of individual neurons
• Interesting mechanism to generate bursting
Reproducing the model
Problems and more problems…
• [Reproducing Polychronization: A
Guide to Maximizing the
Reproducibility of Spiking Network
Models, Pauli et al, 2018]
• Reproduced the original results
with nest::…
• Long story short: not easy, even
with code!
nest::
• NEST is a simulator for spiking neural network models that
focuses on the dynamics, size and structure of neural systems
rather than on the exact morphology of individual neurons
• The development of NEST is coordinated by the NEST Initiative
• NEST is ideal for networks of spiking neurons of any size, for
example:
• Models of information processing e.g. in the visual or auditory cortex
of mammals
• Models of network activity dynamics, e.g. laminar cortical networks or
balanced random networks
• Models of learning and plasticity
Step 0
• Try nest:: desktop online
• Needs EBRAINS account
• Fiddle with setting and parms…
• Only able to use NESTML models
• Default
• From repository
• Conclusion:
• Cool
• Limited (NESTML
required)…
First step
• Install nest::
• From Conda…
• Manually
• From an yml file
• Docker…
• Local Collab…
• Jupyter Notebook!
• Learn a little bit of (pretty easy):
• nest::
• nestml
Next step: Izhikevich nest:: tutorial
• Try to reproduce the
original paper behaviours
• Had to deal with some
nest:: idiosyncrasies
• Sk, not S…
• No reloading NESTML
• Kill & reload kernel!
• No resetting rule
• nest.ResetKernel()
• Rebuild network for every
experiment!
Next: Reproduce single neuron
• No stop until spike-perfect…

v w
Then… 2 ~ 10 neurons!
• From code…

• Spikes/Deltas/Convolutions…?
Ask the forums!
• Not much
action… :-(
• ~ 1 week
• ~ 1 answer
• Not always
useful
Personal conclusions…
• A full month and still
going on…
• Very elegant platform
• New gold standard?
• Specially if planning to run
thousands of neurons
• GPU/Cluster cappabilities
• However…
• Documentation not perfect (spikes!)
• Forums not very active, but they answer in the end, sort of…
• With my own code, ony a couple of weeks, probably
• Not worth the effort… :-(
Final piece of advice
• Stay safe and far from the water!
CNS, JC talks
Exact mean-field models for
spiking neural networks with
adaptation
Liang Chen, Sue Ann Campbell
Journal of Computational Neuroscience (2022) 50:445–469
https://doi.org/10.1007/s10827-022-00825-9

Thanks!

You might also like